jwfromm commented on issue #4464: [RFC] Add TVMDSOOp to integrate any TVM operator with TensorFlow URL: https://github.com/apache/incubator-tvm/issues/4464#issuecomment-562681167 The motivations of this RFC are extremely similar to those in [pytorch-tvm](https://github.com/pytorch/tvm), however the two implementations are very different and it is worth discussing the tradeoffs. - torch-tvm is self contained, it doesn't use any special functions or classes in TVM. Instead it modifies torch script to use existing TVM functions. - torch-tvm uses relay to represent subgraphs and then dynamically builds functions rather than using prebuilt libraries as proposed here. I understand that the current implementation is the shortest path to getting tvm functions working in TensorFlow and that a torch-tvm approach would be a much larger undertaking. However, I don't think it will be able to scale well. The use of prebuilt libraries means there will be a lot of back and forth between regular tvm and tensorflow-tvm during development, and it seems like developers would be better off just importing their tf model to relay and doing everything within tvm. Contrast this to the torch-tvm approach where all the tvm magic happens transparently, making it very straight forward for pytorch users. We should also consider where the code belongs. I personally prefer having projects like torch-tvm and tf-tvm being separate from the main tvm repo if possible as it we already are dealing with frontend bloat. All that said, I think something like tf-tvm is a great idea and something we should work towards. I just want to make sure we make the first step carefully.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
